Research on Gear Decision Method Based on Improved Gated Recurrent Unit

Recurrent Neural Network (RNN) play a pivotal role in extracting temporal information from data, serving as an ideal neural network architecture for transforming a driver’s manual shifting experience into automated gear decisions. However, it confronts the ongoing challenge of inadequately addressing long-term sequence dependencies. Despite improvements like Gated Recurrent Unit (GRU) mitigating gradient explosion, the lingering issue of gradient vanishing remains incompletely resolved. In this paper, a method is proposed based on an improved GRU to transform the driver’s manual shifting experience into automatic gear decisions. Firstly, a longitudinal motion model for the vehicle is established, and vehicle driving data are obtained. Secondly, a self-attention mechanism is introduced to enhance the GRU neural network for learning the driver’s manual shifting behavior. Finally, the gear decisions made by the improved GRU are compared with those made by the driver during manual shifting. The comparative results demonstrate that the improved GRU mitigates the issue of inaccurate gear decisions stemming from gradient vanishing. It effectively translates the manual shifting habits of drivers into precise automatic gear-shifting determinations.